Remove the AutoReshape step for images guaranteed to be in a certain size and type

[DavidZSchneider]

In my data set all images will have the same dimensions: 640x640x3.
All I need to model to support is input of Nx3x640x640 of Pytorch tensor.

How can I skip the AutoReshape step and speed up a bit the preprcoessing of the net?

[glenn-jocher]

@DavidZSchneider hello! If all your images are guaranteed to be 640x640x3, you can streamline the preprocessing by modifying the data pipeline to bypass the auto reshape step. This involves adjusting the code where the images are being prepared for the model input. You'll need to locate the section in the dataset or dataloader where the resizing is called and ensure that it's either removed or conditioned to not execute if the images already meet the required dimensions.

Keep in mind that this change is specific to your use case and dataset, and you should ensure that any future datasets you might use also conform to these dimensions or reintroduce the resizing step as needed. Happy coding! 😊